Researchers at the Korea Advanced Institute of Science and Technology (KAIST) developed a technique that can cut additional training time for AI models by up to 69.6 percent while largely maintaining performance. KAIST said the method, called CluSTER, selects data with similar training effects and distributes it efficiently across multiple graphics processing units. Tests using coding and medical datasets showed the approach maintained performance using about 31 percent of data.

AI chatbot services that virtually recreate conversations with real idols are drawing attention, along with the technologies behind them. The core is upgrading persona techniques used to improve large language model responses by combining them with information about real people and long-term memory. Mindlogic’s fan platform BloomingTalk enables chats and voice or video calls with AI avatars of 24 members of A2O Soul and Light. The company automates persona collection, structures key user data and prioritises reliable inputs.